But also experiment with recordsizes of 8k and 16k. The larger the recordsize, the more potential compression you get, which may be offset by additional delay caused by read/modify/write of the larger record (i.e. reading/writing 32k instead of 8k or 16k).